Research On Ground Truth Analysis Based On Variational Inference

With the fast growth of online data,more and more applications require inferring ground truth from noisy data,especially in wireless sensor networks(WSNs),crowdsourcing networks and recommendation networks.Variational inference(VI)is an important type of ground truth analysis models that can be used to extract information from the noisy data.However,there are some limitations in the existing literature,mainly including sub-optimality and too specific modeling.Furthermore,VI's accuracy and complexity rely heavily on the complexity of the approximate variational distribution,and VI models need to be constructed for specific cases.Therefore,it remains a challenge to find more accurate and general VI models that are applicable for different scenarios.Therefore,this thesis develops the variations and applications of variational inference to solve various ground truth analysis problems.Specifically,we explore the combination of variational inference with message passing algorithm,tempering and Stein identity,and apply them in crowdsourcing inference,channel estimation and recommendation systems.These algorithms are simple and efficient,and significantly advance our ability to solve these challenging problems.We show that our algorithms significantly outperform existing approaches in terms of both empirical performance and theoretical properties.The main outcomes and contributions of our research are the following:1.A variational inference based ground truth analysis method in crowdsourcing networks.In this work,we propose a novel crowdsourcing inference algorithm to infer ground truth and obtain worker reliability and task difficulty at the same time.We first formulate a novel variational message passing algorithm(NMP)to compute task ground truth efficiently.Then we analyze the performance and convergence of NMP both theoretically and experimentally.Finally,we demonstrate that the accuracy of our algorithm is better than others.2.A variational inference with tempering based ground truth analysis method in crowdsourcing networks.In this work,we develop a variational tempering inference(VTI)algorithm and it has been empirically validated and theoretically justified.The numerical experiments of the real-world data demonstrate that our variational tempering inference algorithm performs better than the existing ground truth analysis algorithms.3.A variational inference with tempering based ground truth analysis method in wireless sensor networks(CEVTI).In this work,we consider the channel estimation problem of the pilot signal and channel coefficients,assuming there is orthogonal access between different sensors(or users)and the data fusion center(or receiving center in the following).By formulating the channel estimation problem into a probabilistic graphical model,the proposed CEVTI approach can estimate the channel coefficient and transmit signal in a low complexity manner and guarantee convergence at the same time.Experiments and simulations demonstrate that CEVTI has higher accuracy than state-of-the-art ground truth analysis algorithms.4.A Stein variational inference based ground truth analysis method in recommendation networks.In this work,we propose a more general SVRS to tackle the long plaguing recommendation problem.The SVRS algorithm is based on Stein's identity,and has the merits of low complexity,easy to scale and generalize.Experiments demonstrate that SVRS has higher accuracy in terms of mean average error(MAE)and root mean square error(RMSE).